What Was Prespecified and What Appears to Have Been Decided Later?

Which decisions existed before the results could influence them?

A published paper presents a polished account of a study. The research question appears settled, the outcomes look inevitable, the analysis seems orderly, and the final tables fit together rather nicely. Real research is rarely that linear.

Objectives can change. Outcomes can be added, removed, or reclassified. Eligibility criteria may be modified. Statistical models can change. Subgroups may become interesting only after patterns appear. Some changes are entirely legitimate, but you need to know what was planned before relevant results were known and what was decided later if that distinction affects how strongly you interpret the evidence.

Compare the final paper with documentation created earlier

Reconstruct the chronology of important research decisions

Prespecification is about timing, not merely documentation

A decision is meaningfully prespecified when it is documented before information that could improperly influence that decision becomes available.

The exact relevant time depends on the decision and study design. In a randomized trial, an analysis may ideally be specified before investigators have access to unblinded comparative outcome information. A hypothesis documented after investigators have already examined the relationship in the dataset is not rendered prespecified simply because it later appears in a formal document.

The practical question is therefore not just “Was this written somewhere?” but “When was it written relative to access to information that could influence the choice?”

The publication itself cannot prove prespecification

A methods section may say that an analysis was “planned” or “prespecified.” That statement is useful, but independent dated documentation provides stronger evidence of what existed before the results were known.

For randomized trials, CONSORT 2025 requires access information for the trial protocol and statistical analysis plan and asks authors to report important changes after trial commencement, including outcomes or analyses that were not prespecified.

When the chronology matters, this is why you may need to consult a protocol, registry entry, supplement, or previous paper rather than relying solely on the final publication.

What kinds of decisions can be prespecified?

Prespecification extends well beyond choosing one statistical test.

Not every study needs all of these decisions fixed rigidly in advance. Exploratory research legitimately allows more analytical development. What matters is that the eventual claims match the role and timing of the decisions that generated them.

The protocol, registration, and statistical analysis plan serve different purposes

A protocol describes the planned study in broad methodological detail. A registration record provides a public record of specified study information and can establish dates and versions. A statistical analysis plan can specify analytical procedures in considerably greater detail.

These documents may overlap, but they are not interchangeable. The registration may contain only a concise outcome description while the protocol defines the intervention and design more fully. The statistical analysis plan may specify models, covariates, missing-data procedures, subgroup analyses, and sensitivity analyses that the protocol mentions only generally.

Use the document appropriate to the decision you are investigating.

Compare versions, not merely documents

Protocols and registrations can be amended. That is not inherently problematic. Research sometimes must adapt to recruitment problems, external evidence, operational constraints, measurement failures, safety concerns, or other developments.

CONSORT 2025 explicitly recognizes that changes can occur after a trial begins and asks authors to report the nature, timing, and reasons for important changes. The timing matters because a change made before versus after unblinded comparative information becomes available can carry different implications for bias.

Therefore, do not simply ask whether a protocol exists. Check which version predates the relevant decision and whether amendments are dated.

Outcome changes deserve particular attention

Researchers may add an outcome, remove one, change its definition, alter the time point, convert a secondary outcome into a primary one, or reverse the hierarchy.

Some changes can be justified. Perhaps an instrument becomes unavailable, an external development changes clinical relevance, or a measurement proves impossible before outcome data are examined.

The concern arises when changes are undisclosed or appear related to observed results. CONSORT 2025 notes evidence of discrepancies between outcomes specified in protocols or registries and those reported in final publications, often favoring statistically significant findings, and requires changes to be reported with reasons.

Remember that changing the outcome definition can mean changing the question itself.

Analytical changes can be subtler than outcome changes

The outcome may remain identical while the model changes. Researchers might alter covariates, transformation rules, exclusion criteria, missing-data procedures, subgroup cutoffs, treatment of outliers, time windows, or the definition of the analysis population.

Some flexibility is unavoidable, particularly when unanticipated data characteristics emerge. But these decisions can affect estimates and uncertainty. The more consequential the decision, the more useful it is to know whether it was specified before the relevant results were available.

Prespecified does not mean methodologically sound

A poorly chosen analysis does not become statistically appropriate because it was written into a protocol six months earlier.

Prespecification addresses one methodological concern: the possibility that analytical choices were influenced by observed data or results. It does not guarantee good measurement, correct modeling, appropriate adjustment, adequate power, or freedom from bias.

You generally want both, but they answer different appraisal questions.

Post hoc does not mean fraudulent or useless

Later decisions can be necessary and scientifically informative. A previously unknown data problem may require a new analysis. An unexpected pattern may generate a worthwhile hypothesis. A reviewer may request an additional model that clarifies interpretation.

The problem is not that the analysis happened later. The problem is concealment of its timing or interpretation that treats a data-generated hypothesis as though it had been independently confirmed by the same data.

Transparent labeling lets readers interpret the evidence accordingly.

Subgroup definitions are particularly vulnerable to data-driven flexibility

Suppose age was originally intended to be examined continuously, but the final paper reports effects separately for participants younger and older than 47 years. Why 47?

If that threshold was chosen because it produced the clearest difference, the subgroup analysis has a different evidential status from one based on a clinically justified threshold specified before analysis.

CONSORT 2025 warns against selecting cut points based on statistical significance and asks authors to state the rationale for subgroup definitions and whether analyses were specified a priori or performed post hoc.

Sensitivity analyses should also be planned around identifiable assumptions

ICH E9(R1) recommends prespecifying sensitivity analyses that investigate assumptions underlying the main estimator and cautions against changing many aspects of the analysis simultaneously.

This does not mean that every useful sensitivity analysis must have been anticipated. Unexpected problems may justify additional analyses. But when such analyses are introduced later, their timing and rationale should remain visible.

Exploratory research requires a different expectation

Not all research is confirmatory. Early-stage, descriptive, qualitative, computational, and hypothesis-generating studies may intentionally allow questions and analytical strategies to develop as researchers learn from the data.

Prespecification should therefore not be turned into a universal ritual imposed identically on every design. The more confirmatory the claim, however, the more consequential it becomes to distinguish decisions made independently of the observed results from those developed after seeing them.

Sometimes you cannot determine what was prespecified

A paper may have no accessible protocol, registration, or dated analysis plan. The authors may describe an analysis as planned without providing documentation that allows its timing to be verified.

Do not automatically conclude that the analysis was post hoc. Equally, do not treat unverified prespecification as established fact. Record what the available documentation allows you to know.

Reconstructing how an apparently primary finding became primary

Common mistakes when judging prespecification

Build a timeline rather than assigning a suspicious-or-safe label

For consequential analytical decisions, record the earliest dated source in which you can find them. Then compare later versions and the final publication.

This chronological approach is more useful than treating prespecification as a checkbox. It allows legitimate methodological adaptation while making data-dependent flexibility visible when it matters for interpretation.

Before calling an analysis prespecified, reconstruct when the decisions were made

Questions about prespecified and later research decisions

Prespecification is best understood as a chronology of decisions

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.